Digital Stereo Image Matching Techniques

This paper presents an overview of the various classes of algorithms in use for matching points from one digital image of a stereo pair with the corresponding points in the second image of the pair. These techniques primarily use area-based measures, such as correlation between image patches, or edge-based methods that match linear features in images, but also include the use of feature extractors to match single points in images, as well as global optimization techniques that simultaneously match all points in the two images. This paper also describes an automatic system developed at SRI for stereo compilation; this system uses area-based correlation , but applies this basic technique in a variety of novel ways to develop a disparity model for a given stereo image pair. The techniques used are hierarchical in nature, and incorporate iterative refinement, as well as a best-first strategy, in the matching process. To illustrate these techniques, the results of this system on the Image Matching Test A data set recently distributed by ISPRS's Working Group III/4 are presented. Introd uction Automatic techniques for the production of three-dimensional (3-D) data via stereo conlpila-tion are receiving increased interest for a variety of applications, including cartography [Panton, 1978], autonomous vehicle navigation [Hannah, 1980], and industrial automation [Nishihara and Poggio, 1983]. The first and most difficult step in recovering 3-D information frolll a pair of stereo images is that of matching points from one digital image of the pair with the corresponding points in the second image. Many computational algorithms have been used in attempts to solve this problem (see [Brady, 1982] or [Barnard and Fischler, 1982] for surveys of the field). These techniques primarily use area-based measures, such as correlation between image patches, or edge-based methods that match linear features in images, but also include the use of feature extractors to match single points in images, as well as global optimization techniques that simultaneously match all points .in the two images. Area matching techniques are the oldest and simplest of the stereo matching algorithms. Each image point to be matched is in fact the center of a small window of points' in the first or reference image, and this window is statistically compared with similarly sized windows of points in the second or target image of the stereo pair. The measure of match is either a difference metric that is minimized, such as RMS difference, or more commonly …